=Paper=
{{Paper
|id=Vol-2098/abstract5
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2098/abstract5.pdf
|volume=Vol-2098
}}
==None==
<pdf width="1500px">https://ceur-ws.org/Vol-2098/abstract5.pdf</pdf>
<pre>
             Modern Nonconvex Optimization.

            Theory, Methods, and Application


                            Alexander S. Strekalovsky
 Matrosov Institute for System Dynamics and Control Theory of SB RAS, Irkutsk,

                              Russia   strekal@icc.ru


Keywords:     nonlinear optimization, d.c. optimization, penalized problem, local
search, global search.

    We consider the general nonconvex problem with the goal function and
equality and inequality constraints given by d.c. functions.
    We reduce this problem to a problem without nonconvex constraints by the
exact penalty approach and present the penalized problem as a d.c. minimization
one, i.e. with d.c. goal function which is nonsmooth. Furthermore, relations
between the original problem and the penalized problem are investigated.
    In addition, an employment the d.c. structure of penalized problem the
Global Optimality Conditions (GOCs) are developed and analysed. In particular,
we prove that the GOCs possess the constructive property, i.e. when the GOCs
are violated, it is possible to nd a feasible (in original problem) vector which is
better than the point under investigation.
    Moreover, it is shown that the point satisfying the GOCs turns out to be a
KKT vector in the original problem. It means, that the new GOCs are related
to the Classical Optimization Theory.
    Besides we establish that the verication of the GOCs consists in a solution
of a family of the partially linearized (with respect to the basic nonconvexities
of the original problem unied by the exact penalty in one function) problem,
and consecutive verication of the principal inequality of the GOCs.
    The eectiveness of the GOCs is veried by a number of examples in which
the GOCs conrm its ability to escape stationary points and local minima with
improving the goal function.
    On the base of GOC we propose local and global search methods for the
penalized problems and investigate their convergence and other properties.
    The testing of the developed Global Search Theory were carried out on a
rather large eld of test examples and applied problems of dierent nature. For
instance, there were considered quadratic and polynomial extremum problems
and fractional optimization problems.
    In addition, a seeking for a solution to d.c. equations systems, the search for
the Nash equilibrium points, the nding of optimistic and pessimistic solutions
to bilevel problems with (d.c.) nonlinear data have been performed.
    The results of computational simulations look rather promising even for the
case of problems of high dimensions.

</pre>